Multi-spectral semantic segmentation and localization framework for ecosystem monitoring using Visible–SWIR satellite imagery

Abstract

This study proposes a Visible–SWIR multispectral semantic segmentation and geospatial localization framework for ecosystem monitoring. The method fuses Visible, SWIR1, SWIR2, NDVI, NBR, NBR2, NDMI, and brightness temperature features to enhance the identification of vegetation, moisture, burned surfaces, and thermal anomalies. The segmentation outputs are converted into object-level descriptors—class, area, centroid, distance, and orientation—and paired with satellite metadata and coordinate constraints to enable localization. In the Kangaroo Island wildfire case study, the proposed method correctly detected 5,295 fire pixels, achieving 90.51% precision, 80.32% recall, an F1 score of 85.11%, and an IoU of 74.09%. Results show that multispectral fusion improves fire-core recognition and boundary continuity while mitigating single-band misclassifications caused by smoke, clouds, and surface heterogeneity. The proposed framework can be further applied to land-cover evaluation, vegetation monitoring, disaster assessment, and time-series ecological change detection.